期刊
JOURNAL OF CLEANER PRODUCTION
卷 216, 期 -, 页码 288-310出版社
ELSEVIER SCI LTD
DOI: 10.1016/j.jclepro.2019.01.158
关键词
ECMWF-Based solar prediction model; Temperature models; Machine learning models; Neural networks; Feature selection
资金
- Research and Training Scheme [RTS 2016-2021]
To support alternative forms of energy resources, the prediction of global incident solar radiation (Irad) is critical to establish the efficacy of solar energy resources as a free and clean energy, and to identify and screen solar powered sites. Solar radiation data for construction of energy feasibility studies are not available in many locations due to the absence of meteorological stations, especially in remote or regional sites. To surmount the challenge in solar energy site identification, the universally gridded data integrated into predictive models used to generate reliable Irad forecasts can be considered as a viable medium for future energy utilization. The objective of this paper is to review, develop and evaluate a suite of machine learning (ML) models based on the artificial neural network (ANN) versus several other kinds of data-driven models such as support vector regression (SVR), Gaussian process machine learning (GPML) and genetic programming (GP) models for the prediction of daily Irad generated through the European Centre for Medium Range Weather Forecasting (ECMWF) Reanalysis fields. The performance of the ML models are benchmarked against several statistical tools: auto regressive moving integrated average (ARIMA), Temperature Model (TM), Time series and Fourier Series (TSFS) models. To train these models, 87 different predictor variables from the ERA-Interim reanalysis dataset (01-January-1979 to 31 December -2015) were extracted for 5 solar-rich metropolitan sites (i.e., Brisbane, Gold Coast, Sunshine Coast, Ipswich and Toowoomba, Australia) targeted against surface level Irad available from the measured Scientific Information for Land Owners dataset. For daily forecast models, a total of the 20 most important predictors related to the Irad dataset were screened with nearest component analysis: fsrnca feature selection, and partitioned into training (70%), validation (15%) and testing (15%) sets for model design. To benchmark the ANN, TSFS and TM models were developed with Fourier series and regression analysis, respectively and the statistical performance was benchmarked with root mean square error (RMSE), mean absolute error (MAE), Nash-Sutcliffe efficiency (ENS), Willmott's Index (WI), Mean Bias error (MBE), Legates and McCabe Index (El), and relative MAE, RMSE and diagnostic plots. The performance of ANN was significantly better than the other models (SVR, GPML, GP, TM), resulting in lower RMSE (1.715 2.27 MJM-2/day relative to 2.14-5.90 MJrn-2/day), relative RMSE (9.07-12.47 vs 10.98-29.15), relative RMAE (7.97-11.74 vs 9.27-33.96) and larger WI, ENS and E1 (0.938-0.967 vs. 0.462-0.955, 0.935-0.872 vs. 0.355-0.915, 0.672-0.783 vs. 0.252-0.740). Additionally, models assessed with predictors grouped into El Nino, La Nina and the positive, negative and neutral periods of Indian Ocean Dipole, affirmed the merits of ANN model (RRMSE < 11%). Seasonal analysis showed that ANN was an elite tool over SVR. GPML and GP for Irad prediction. The study concludes that an ANN approach integrated with ECMWF fields, incorporating physical interactions of Ira with atmospheric data, is an efficacious alternative to forecast solar energy and assist with energy modelling for solar-rich sites that have diverse climatic conditions to further support clean energy utilization. (C) 2019 Elsevier Ltd. All rights reserved.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据